1 // SPDX-License-Identifier: GPL-2.0
2 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
5 #include <linux/sched.h>
6 #include <linux/sched/mm.h>
7 #include <linux/sched/coredump.h>
8 #include <linux/mmu_notifier.h>
9 #include <linux/rmap.h>
10 #include <linux/swap.h>
11 #include <linux/mm_inline.h>
12 #include <linux/kthread.h>
13 #include <linux/khugepaged.h>
14 #include <linux/freezer.h>
15 #include <linux/mman.h>
16 #include <linux/hashtable.h>
17 #include <linux/userfaultfd_k.h>
18 #include <linux/page_idle.h>
19 #include <linux/swapops.h>
20 #include <linux/shmem_fs.h>
23 #include <asm/pgalloc.h>
32 SCAN_EXCEED_SHARED_PTE,
36 SCAN_LACK_REFERENCED_PAGE,
50 SCAN_ALLOC_HUGE_PAGE_FAIL,
51 SCAN_CGROUP_CHARGE_FAIL,
53 SCAN_PAGE_HAS_PRIVATE,
56 #define CREATE_TRACE_POINTS
57 #include <trace/events/huge_memory.h>
59 static struct task_struct *khugepaged_thread __read_mostly;
60 static DEFINE_MUTEX(khugepaged_mutex);
62 /* default scan 8*512 pte (or vmas) every 30 second */
63 static unsigned int khugepaged_pages_to_scan __read_mostly;
64 static unsigned int khugepaged_pages_collapsed;
65 static unsigned int khugepaged_full_scans;
66 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
67 /* during fragmentation poll the hugepage allocator once every minute */
68 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
69 static unsigned long khugepaged_sleep_expire;
70 static DEFINE_SPINLOCK(khugepaged_mm_lock);
71 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
73 * default collapse hugepages if there is at least one pte mapped like
74 * it would have happened if the vma was large enough during page
77 static unsigned int khugepaged_max_ptes_none __read_mostly;
78 static unsigned int khugepaged_max_ptes_swap __read_mostly;
79 static unsigned int khugepaged_max_ptes_shared __read_mostly;
81 #define MM_SLOTS_HASH_BITS 10
82 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
84 static struct kmem_cache *mm_slot_cache __read_mostly;
86 #define MAX_PTE_MAPPED_THP 8
89 * struct mm_slot - hash lookup from mm to mm_slot
90 * @hash: hash collision list
91 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
92 * @mm: the mm that this information is valid for
95 struct hlist_node hash;
96 struct list_head mm_node;
99 /* pte-mapped THP in this mm */
100 int nr_pte_mapped_thp;
101 unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
105 * struct khugepaged_scan - cursor for scanning
106 * @mm_head: the head of the mm list to scan
107 * @mm_slot: the current mm_slot we are scanning
108 * @address: the next address inside that to be scanned
110 * There is only the one khugepaged_scan instance of this cursor structure.
112 struct khugepaged_scan {
113 struct list_head mm_head;
114 struct mm_slot *mm_slot;
115 unsigned long address;
118 static struct khugepaged_scan khugepaged_scan = {
119 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
123 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
124 struct kobj_attribute *attr,
127 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
130 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
131 struct kobj_attribute *attr,
132 const char *buf, size_t count)
137 err = kstrtoul(buf, 10, &msecs);
138 if (err || msecs > UINT_MAX)
141 khugepaged_scan_sleep_millisecs = msecs;
142 khugepaged_sleep_expire = 0;
143 wake_up_interruptible(&khugepaged_wait);
147 static struct kobj_attribute scan_sleep_millisecs_attr =
148 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
149 scan_sleep_millisecs_store);
151 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
152 struct kobj_attribute *attr,
155 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
158 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
159 struct kobj_attribute *attr,
160 const char *buf, size_t count)
165 err = kstrtoul(buf, 10, &msecs);
166 if (err || msecs > UINT_MAX)
169 khugepaged_alloc_sleep_millisecs = msecs;
170 khugepaged_sleep_expire = 0;
171 wake_up_interruptible(&khugepaged_wait);
175 static struct kobj_attribute alloc_sleep_millisecs_attr =
176 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
177 alloc_sleep_millisecs_store);
179 static ssize_t pages_to_scan_show(struct kobject *kobj,
180 struct kobj_attribute *attr,
183 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
185 static ssize_t pages_to_scan_store(struct kobject *kobj,
186 struct kobj_attribute *attr,
187 const char *buf, size_t count)
192 err = kstrtoul(buf, 10, &pages);
193 if (err || !pages || pages > UINT_MAX)
196 khugepaged_pages_to_scan = pages;
200 static struct kobj_attribute pages_to_scan_attr =
201 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
202 pages_to_scan_store);
204 static ssize_t pages_collapsed_show(struct kobject *kobj,
205 struct kobj_attribute *attr,
208 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
210 static struct kobj_attribute pages_collapsed_attr =
211 __ATTR_RO(pages_collapsed);
213 static ssize_t full_scans_show(struct kobject *kobj,
214 struct kobj_attribute *attr,
217 return sprintf(buf, "%u\n", khugepaged_full_scans);
219 static struct kobj_attribute full_scans_attr =
220 __ATTR_RO(full_scans);
222 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
223 struct kobj_attribute *attr, char *buf)
225 return single_hugepage_flag_show(kobj, attr, buf,
226 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
228 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
229 struct kobj_attribute *attr,
230 const char *buf, size_t count)
232 return single_hugepage_flag_store(kobj, attr, buf, count,
233 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
235 static struct kobj_attribute khugepaged_defrag_attr =
236 __ATTR(defrag, 0644, khugepaged_defrag_show,
237 khugepaged_defrag_store);
240 * max_ptes_none controls if khugepaged should collapse hugepages over
241 * any unmapped ptes in turn potentially increasing the memory
242 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
243 * reduce the available free memory in the system as it
244 * runs. Increasing max_ptes_none will instead potentially reduce the
245 * free memory in the system during the khugepaged scan.
247 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
248 struct kobj_attribute *attr,
251 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
253 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
254 struct kobj_attribute *attr,
255 const char *buf, size_t count)
258 unsigned long max_ptes_none;
260 err = kstrtoul(buf, 10, &max_ptes_none);
261 if (err || max_ptes_none > HPAGE_PMD_NR-1)
264 khugepaged_max_ptes_none = max_ptes_none;
268 static struct kobj_attribute khugepaged_max_ptes_none_attr =
269 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
270 khugepaged_max_ptes_none_store);
272 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
273 struct kobj_attribute *attr,
276 return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
279 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
280 struct kobj_attribute *attr,
281 const char *buf, size_t count)
284 unsigned long max_ptes_swap;
286 err = kstrtoul(buf, 10, &max_ptes_swap);
287 if (err || max_ptes_swap > HPAGE_PMD_NR-1)
290 khugepaged_max_ptes_swap = max_ptes_swap;
295 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
296 __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
297 khugepaged_max_ptes_swap_store);
299 static ssize_t khugepaged_max_ptes_shared_show(struct kobject *kobj,
300 struct kobj_attribute *attr,
303 return sprintf(buf, "%u\n", khugepaged_max_ptes_shared);
306 static ssize_t khugepaged_max_ptes_shared_store(struct kobject *kobj,
307 struct kobj_attribute *attr,
308 const char *buf, size_t count)
311 unsigned long max_ptes_shared;
313 err = kstrtoul(buf, 10, &max_ptes_shared);
314 if (err || max_ptes_shared > HPAGE_PMD_NR-1)
317 khugepaged_max_ptes_shared = max_ptes_shared;
322 static struct kobj_attribute khugepaged_max_ptes_shared_attr =
323 __ATTR(max_ptes_shared, 0644, khugepaged_max_ptes_shared_show,
324 khugepaged_max_ptes_shared_store);
326 static struct attribute *khugepaged_attr[] = {
327 &khugepaged_defrag_attr.attr,
328 &khugepaged_max_ptes_none_attr.attr,
329 &khugepaged_max_ptes_swap_attr.attr,
330 &khugepaged_max_ptes_shared_attr.attr,
331 &pages_to_scan_attr.attr,
332 &pages_collapsed_attr.attr,
333 &full_scans_attr.attr,
334 &scan_sleep_millisecs_attr.attr,
335 &alloc_sleep_millisecs_attr.attr,
339 struct attribute_group khugepaged_attr_group = {
340 .attrs = khugepaged_attr,
341 .name = "khugepaged",
343 #endif /* CONFIG_SYSFS */
345 int hugepage_madvise(struct vm_area_struct *vma,
346 unsigned long *vm_flags, int advice)
352 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
353 * can't handle this properly after s390_enable_sie, so we simply
354 * ignore the madvise to prevent qemu from causing a SIGSEGV.
356 if (mm_has_pgste(vma->vm_mm))
359 *vm_flags &= ~VM_NOHUGEPAGE;
360 *vm_flags |= VM_HUGEPAGE;
362 * If the vma become good for khugepaged to scan,
363 * register it here without waiting a page fault that
364 * may not happen any time soon.
366 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
367 khugepaged_enter_vma_merge(vma, *vm_flags))
370 case MADV_NOHUGEPAGE:
371 *vm_flags &= ~VM_HUGEPAGE;
372 *vm_flags |= VM_NOHUGEPAGE;
374 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
375 * this vma even if we leave the mm registered in khugepaged if
376 * it got registered before VM_NOHUGEPAGE was set.
384 int __init khugepaged_init(void)
386 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
387 sizeof(struct mm_slot),
388 __alignof__(struct mm_slot), 0, NULL);
392 khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
393 khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
394 khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
395 khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
400 void __init khugepaged_destroy(void)
402 kmem_cache_destroy(mm_slot_cache);
405 static inline struct mm_slot *alloc_mm_slot(void)
407 if (!mm_slot_cache) /* initialization failed */
409 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
412 static inline void free_mm_slot(struct mm_slot *mm_slot)
414 kmem_cache_free(mm_slot_cache, mm_slot);
417 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
419 struct mm_slot *mm_slot;
421 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
422 if (mm == mm_slot->mm)
428 static void insert_to_mm_slots_hash(struct mm_struct *mm,
429 struct mm_slot *mm_slot)
432 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
435 static inline int khugepaged_test_exit(struct mm_struct *mm)
437 return atomic_read(&mm->mm_users) == 0;
440 static bool hugepage_vma_check(struct vm_area_struct *vma,
441 unsigned long vm_flags)
443 if (!transhuge_vma_enabled(vma, vm_flags))
446 if (vma->vm_file && !IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) -
447 vma->vm_pgoff, HPAGE_PMD_NR))
450 /* Enabled via shmem mount options or sysfs settings. */
451 if (shmem_file(vma->vm_file))
452 return shmem_huge_enabled(vma);
454 /* THP settings require madvise. */
455 if (!(vm_flags & VM_HUGEPAGE) && !khugepaged_always())
458 /* Only regular file is valid */
459 if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && vma->vm_file &&
460 (vm_flags & VM_DENYWRITE)) {
461 struct inode *inode = vma->vm_file->f_inode;
463 return S_ISREG(inode->i_mode);
466 if (!vma->anon_vma || vma->vm_ops)
468 if (vma_is_temporary_stack(vma))
470 return !(vm_flags & VM_NO_KHUGEPAGED);
473 int __khugepaged_enter(struct mm_struct *mm)
475 struct mm_slot *mm_slot;
478 mm_slot = alloc_mm_slot();
482 /* __khugepaged_exit() must not run from under us */
483 VM_BUG_ON_MM(atomic_read(&mm->mm_users) == 0, mm);
484 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
485 free_mm_slot(mm_slot);
489 spin_lock(&khugepaged_mm_lock);
490 insert_to_mm_slots_hash(mm, mm_slot);
492 * Insert just behind the scanning cursor, to let the area settle
495 wakeup = list_empty(&khugepaged_scan.mm_head);
496 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
497 spin_unlock(&khugepaged_mm_lock);
501 wake_up_interruptible(&khugepaged_wait);
506 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
507 unsigned long vm_flags)
509 unsigned long hstart, hend;
512 * khugepaged only supports read-only files for non-shmem files.
513 * khugepaged does not yet work on special mappings. And
514 * file-private shmem THP is not supported.
516 if (!hugepage_vma_check(vma, vm_flags))
519 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
520 hend = vma->vm_end & HPAGE_PMD_MASK;
522 return khugepaged_enter(vma, vm_flags);
526 void __khugepaged_exit(struct mm_struct *mm)
528 struct mm_slot *mm_slot;
531 spin_lock(&khugepaged_mm_lock);
532 mm_slot = get_mm_slot(mm);
533 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
534 hash_del(&mm_slot->hash);
535 list_del(&mm_slot->mm_node);
538 spin_unlock(&khugepaged_mm_lock);
541 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
542 free_mm_slot(mm_slot);
544 } else if (mm_slot) {
546 * This is required to serialize against
547 * khugepaged_test_exit() (which is guaranteed to run
548 * under mmap sem read mode). Stop here (after we
549 * return all pagetables will be destroyed) until
550 * khugepaged has finished working on the pagetables
551 * under the mmap_lock.
554 mmap_write_unlock(mm);
558 static void release_pte_page(struct page *page)
560 mod_node_page_state(page_pgdat(page),
561 NR_ISOLATED_ANON + page_is_file_lru(page),
564 putback_lru_page(page);
567 static void release_pte_pages(pte_t *pte, pte_t *_pte,
568 struct list_head *compound_pagelist)
570 struct page *page, *tmp;
572 while (--_pte >= pte) {
573 pte_t pteval = *_pte;
575 page = pte_page(pteval);
576 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
578 release_pte_page(page);
581 list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
582 list_del(&page->lru);
583 release_pte_page(page);
587 static bool is_refcount_suitable(struct page *page)
589 int expected_refcount;
591 expected_refcount = total_mapcount(page);
592 if (PageSwapCache(page))
593 expected_refcount += compound_nr(page);
595 return page_count(page) == expected_refcount;
598 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
599 unsigned long address,
601 struct list_head *compound_pagelist)
603 struct page *page = NULL;
605 int none_or_zero = 0, shared = 0, result = 0, referenced = 0;
606 bool writable = false;
608 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
609 _pte++, address += PAGE_SIZE) {
610 pte_t pteval = *_pte;
611 if (pte_none(pteval) || (pte_present(pteval) &&
612 is_zero_pfn(pte_pfn(pteval)))) {
613 if (!userfaultfd_armed(vma) &&
614 ++none_or_zero <= khugepaged_max_ptes_none) {
617 result = SCAN_EXCEED_NONE_PTE;
621 if (!pte_present(pteval)) {
622 result = SCAN_PTE_NON_PRESENT;
625 page = vm_normal_page(vma, address, pteval);
626 if (unlikely(!page)) {
627 result = SCAN_PAGE_NULL;
631 VM_BUG_ON_PAGE(!PageAnon(page), page);
633 if (page_mapcount(page) > 1 &&
634 ++shared > khugepaged_max_ptes_shared) {
635 result = SCAN_EXCEED_SHARED_PTE;
639 if (PageCompound(page)) {
641 page = compound_head(page);
644 * Check if we have dealt with the compound page
647 list_for_each_entry(p, compound_pagelist, lru) {
654 * We can do it before isolate_lru_page because the
655 * page can't be freed from under us. NOTE: PG_lock
656 * is needed to serialize against split_huge_page
657 * when invoked from the VM.
659 if (!trylock_page(page)) {
660 result = SCAN_PAGE_LOCK;
665 * Check if the page has any GUP (or other external) pins.
667 * The page table that maps the page has been already unlinked
668 * from the page table tree and this process cannot get
669 * an additinal pin on the page.
671 * New pins can come later if the page is shared across fork,
672 * but not from this process. The other process cannot write to
673 * the page, only trigger CoW.
675 if (!is_refcount_suitable(page)) {
677 result = SCAN_PAGE_COUNT;
680 if (!pte_write(pteval) && PageSwapCache(page) &&
681 !reuse_swap_page(page, NULL)) {
683 * Page is in the swap cache and cannot be re-used.
684 * It cannot be collapsed into a THP.
687 result = SCAN_SWAP_CACHE_PAGE;
692 * Isolate the page to avoid collapsing an hugepage
693 * currently in use by the VM.
695 if (isolate_lru_page(page)) {
697 result = SCAN_DEL_PAGE_LRU;
700 mod_node_page_state(page_pgdat(page),
701 NR_ISOLATED_ANON + page_is_file_lru(page),
703 VM_BUG_ON_PAGE(!PageLocked(page), page);
704 VM_BUG_ON_PAGE(PageLRU(page), page);
706 if (PageCompound(page))
707 list_add_tail(&page->lru, compound_pagelist);
709 /* There should be enough young pte to collapse the page */
710 if (pte_young(pteval) ||
711 page_is_young(page) || PageReferenced(page) ||
712 mmu_notifier_test_young(vma->vm_mm, address))
715 if (pte_write(pteval))
719 if (unlikely(!writable)) {
720 result = SCAN_PAGE_RO;
721 } else if (unlikely(!referenced)) {
722 result = SCAN_LACK_REFERENCED_PAGE;
724 result = SCAN_SUCCEED;
725 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
726 referenced, writable, result);
730 release_pte_pages(pte, _pte, compound_pagelist);
731 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
732 referenced, writable, result);
736 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
737 struct vm_area_struct *vma,
738 unsigned long address,
740 struct list_head *compound_pagelist)
742 struct page *src_page, *tmp;
744 for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
745 _pte++, page++, address += PAGE_SIZE) {
746 pte_t pteval = *_pte;
748 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
749 clear_user_highpage(page, address);
750 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
751 if (is_zero_pfn(pte_pfn(pteval))) {
753 * ptl mostly unnecessary.
757 * paravirt calls inside pte_clear here are
760 pte_clear(vma->vm_mm, address, _pte);
764 src_page = pte_page(pteval);
765 copy_user_highpage(page, src_page, address, vma);
766 if (!PageCompound(src_page))
767 release_pte_page(src_page);
769 * ptl mostly unnecessary, but preempt has to
770 * be disabled to update the per-cpu stats
771 * inside page_remove_rmap().
775 * paravirt calls inside pte_clear here are
778 pte_clear(vma->vm_mm, address, _pte);
779 page_remove_rmap(src_page, false);
781 free_page_and_swap_cache(src_page);
785 list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
786 list_del(&src_page->lru);
787 release_pte_page(src_page);
791 static void khugepaged_alloc_sleep(void)
795 add_wait_queue(&khugepaged_wait, &wait);
796 freezable_schedule_timeout_interruptible(
797 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
798 remove_wait_queue(&khugepaged_wait, &wait);
801 static int khugepaged_node_load[MAX_NUMNODES];
803 static bool khugepaged_scan_abort(int nid)
808 * If node_reclaim_mode is disabled, then no extra effort is made to
809 * allocate memory locally.
811 if (!node_reclaim_mode)
814 /* If there is a count for this node already, it must be acceptable */
815 if (khugepaged_node_load[nid])
818 for (i = 0; i < MAX_NUMNODES; i++) {
819 if (!khugepaged_node_load[i])
821 if (node_distance(nid, i) > node_reclaim_distance)
827 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
828 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
830 return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
834 static int khugepaged_find_target_node(void)
836 static int last_khugepaged_target_node = NUMA_NO_NODE;
837 int nid, target_node = 0, max_value = 0;
839 /* find first node with max normal pages hit */
840 for (nid = 0; nid < MAX_NUMNODES; nid++)
841 if (khugepaged_node_load[nid] > max_value) {
842 max_value = khugepaged_node_load[nid];
846 /* do some balance if several nodes have the same hit record */
847 if (target_node <= last_khugepaged_target_node)
848 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
850 if (max_value == khugepaged_node_load[nid]) {
855 last_khugepaged_target_node = target_node;
859 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
861 if (IS_ERR(*hpage)) {
867 khugepaged_alloc_sleep();
877 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
879 VM_BUG_ON_PAGE(*hpage, *hpage);
881 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
882 if (unlikely(!*hpage)) {
883 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
884 *hpage = ERR_PTR(-ENOMEM);
888 prep_transhuge_page(*hpage);
889 count_vm_event(THP_COLLAPSE_ALLOC);
893 static int khugepaged_find_target_node(void)
898 static inline struct page *alloc_khugepaged_hugepage(void)
902 page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
905 prep_transhuge_page(page);
909 static struct page *khugepaged_alloc_hugepage(bool *wait)
914 hpage = alloc_khugepaged_hugepage();
916 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
921 khugepaged_alloc_sleep();
923 count_vm_event(THP_COLLAPSE_ALLOC);
924 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
929 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
932 * If the hpage allocated earlier was briefly exposed in page cache
933 * before collapse_file() failed, it is possible that racing lookups
934 * have not yet completed, and would then be unpleasantly surprised by
935 * finding the hpage reused for the same mapping at a different offset.
936 * Just release the previous allocation if there is any danger of that.
938 if (*hpage && page_count(*hpage) > 1) {
944 *hpage = khugepaged_alloc_hugepage(wait);
946 if (unlikely(!*hpage))
953 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
962 * If mmap_lock temporarily dropped, revalidate vma
963 * before taking mmap_lock.
964 * Return 0 if succeeds, otherwise return none-zero
968 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
969 struct vm_area_struct **vmap)
971 struct vm_area_struct *vma;
972 unsigned long hstart, hend;
974 if (unlikely(khugepaged_test_exit(mm)))
975 return SCAN_ANY_PROCESS;
977 *vmap = vma = find_vma(mm, address);
979 return SCAN_VMA_NULL;
981 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
982 hend = vma->vm_end & HPAGE_PMD_MASK;
983 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
984 return SCAN_ADDRESS_RANGE;
985 if (!hugepage_vma_check(vma, vma->vm_flags))
986 return SCAN_VMA_CHECK;
987 /* Anon VMA expected */
988 if (!vma->anon_vma || vma->vm_ops)
989 return SCAN_VMA_CHECK;
994 * Bring missing pages in from swap, to complete THP collapse.
995 * Only done if khugepaged_scan_pmd believes it is worthwhile.
997 * Called and returns without pte mapped or spinlocks held,
998 * but with mmap_lock held to protect against vma changes.
1001 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
1002 struct vm_area_struct *vma,
1003 unsigned long address, pmd_t *pmd,
1008 struct vm_fault vmf = {
1011 .flags = FAULT_FLAG_ALLOW_RETRY,
1013 .pgoff = linear_page_index(vma, address),
1016 vmf.pte = pte_offset_map(pmd, address);
1017 for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
1018 vmf.pte++, vmf.address += PAGE_SIZE) {
1019 vmf.orig_pte = *vmf.pte;
1020 if (!is_swap_pte(vmf.orig_pte))
1023 ret = do_swap_page(&vmf);
1025 /* do_swap_page returns VM_FAULT_RETRY with released mmap_lock */
1026 if (ret & VM_FAULT_RETRY) {
1028 if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
1029 /* vma is no longer available, don't continue to swapin */
1030 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1033 /* check if the pmd is still valid */
1034 if (mm_find_pmd(mm, address) != pmd) {
1035 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1039 if (ret & VM_FAULT_ERROR) {
1040 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1043 /* pte is unmapped now, we need to map it */
1044 vmf.pte = pte_offset_map(pmd, vmf.address);
1049 /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
1053 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
1057 static void collapse_huge_page(struct mm_struct *mm,
1058 unsigned long address,
1059 struct page **hpage,
1060 int node, int referenced, int unmapped)
1062 LIST_HEAD(compound_pagelist);
1066 struct page *new_page;
1067 spinlock_t *pmd_ptl, *pte_ptl;
1068 int isolated = 0, result = 0;
1069 struct vm_area_struct *vma;
1070 struct mmu_notifier_range range;
1073 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1075 /* Only allocate from the target node */
1076 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1079 * Before allocating the hugepage, release the mmap_lock read lock.
1080 * The allocation can take potentially a long time if it involves
1081 * sync compaction, and we do not need to hold the mmap_lock during
1082 * that. We will recheck the vma after taking it again in write mode.
1084 mmap_read_unlock(mm);
1085 new_page = khugepaged_alloc_page(hpage, gfp, node);
1087 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1091 if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1092 result = SCAN_CGROUP_CHARGE_FAIL;
1095 count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1098 result = hugepage_vma_revalidate(mm, address, &vma);
1100 mmap_read_unlock(mm);
1104 pmd = mm_find_pmd(mm, address);
1106 result = SCAN_PMD_NULL;
1107 mmap_read_unlock(mm);
1112 * __collapse_huge_page_swapin always returns with mmap_lock locked.
1113 * If it fails, we release mmap_lock and jump out_nolock.
1114 * Continuing to collapse causes inconsistency.
1116 if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1118 mmap_read_unlock(mm);
1122 mmap_read_unlock(mm);
1124 * Prevent all access to pagetables with the exception of
1125 * gup_fast later handled by the ptep_clear_flush and the VM
1126 * handled by the anon_vma lock + PG_lock.
1128 mmap_write_lock(mm);
1129 result = hugepage_vma_revalidate(mm, address, &vma);
1132 /* check if the pmd is still valid */
1133 if (mm_find_pmd(mm, address) != pmd)
1136 anon_vma_lock_write(vma->anon_vma);
1138 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1139 address, address + HPAGE_PMD_SIZE);
1140 mmu_notifier_invalidate_range_start(&range);
1142 pte = pte_offset_map(pmd, address);
1143 pte_ptl = pte_lockptr(mm, pmd);
1145 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1147 * After this gup_fast can't run anymore. This also removes
1148 * any huge TLB entry from the CPU so we won't allow
1149 * huge and small TLB entries for the same virtual address
1150 * to avoid the risk of CPU bugs in that area.
1152 _pmd = pmdp_collapse_flush(vma, address, pmd);
1153 spin_unlock(pmd_ptl);
1154 mmu_notifier_invalidate_range_end(&range);
1157 isolated = __collapse_huge_page_isolate(vma, address, pte,
1158 &compound_pagelist);
1159 spin_unlock(pte_ptl);
1161 if (unlikely(!isolated)) {
1164 BUG_ON(!pmd_none(*pmd));
1166 * We can only use set_pmd_at when establishing
1167 * hugepmds and never for establishing regular pmds that
1168 * points to regular pagetables. Use pmd_populate for that
1170 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1171 spin_unlock(pmd_ptl);
1172 anon_vma_unlock_write(vma->anon_vma);
1178 * All pages are isolated and locked so anon_vma rmap
1179 * can't run anymore.
1181 anon_vma_unlock_write(vma->anon_vma);
1183 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1184 &compound_pagelist);
1186 __SetPageUptodate(new_page);
1187 pgtable = pmd_pgtable(_pmd);
1189 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1190 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1193 * spin_lock() below is not the equivalent of smp_wmb(), so
1194 * this is needed to avoid the copy_huge_page writes to become
1195 * visible after the set_pmd_at() write.
1200 BUG_ON(!pmd_none(*pmd));
1201 page_add_new_anon_rmap(new_page, vma, address, true);
1202 lru_cache_add_inactive_or_unevictable(new_page, vma);
1203 pgtable_trans_huge_deposit(mm, pmd, pgtable);
1204 set_pmd_at(mm, address, pmd, _pmd);
1205 update_mmu_cache_pmd(vma, address, pmd);
1206 spin_unlock(pmd_ptl);
1210 khugepaged_pages_collapsed++;
1211 result = SCAN_SUCCEED;
1213 mmap_write_unlock(mm);
1215 if (!IS_ERR_OR_NULL(*hpage))
1216 mem_cgroup_uncharge(*hpage);
1217 trace_mm_collapse_huge_page(mm, isolated, result);
1223 static int khugepaged_scan_pmd(struct mm_struct *mm,
1224 struct vm_area_struct *vma,
1225 unsigned long address,
1226 struct page **hpage)
1230 int ret = 0, result = 0, referenced = 0;
1231 int none_or_zero = 0, shared = 0;
1232 struct page *page = NULL;
1233 unsigned long _address;
1235 int node = NUMA_NO_NODE, unmapped = 0;
1236 bool writable = false;
1238 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1240 pmd = mm_find_pmd(mm, address);
1242 result = SCAN_PMD_NULL;
1246 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1247 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1248 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1249 _pte++, _address += PAGE_SIZE) {
1250 pte_t pteval = *_pte;
1251 if (is_swap_pte(pteval)) {
1252 if (++unmapped <= khugepaged_max_ptes_swap) {
1254 * Always be strict with uffd-wp
1255 * enabled swap entries. Please see
1256 * comment below for pte_uffd_wp().
1258 if (pte_swp_uffd_wp(pteval)) {
1259 result = SCAN_PTE_UFFD_WP;
1264 result = SCAN_EXCEED_SWAP_PTE;
1268 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1269 if (!userfaultfd_armed(vma) &&
1270 ++none_or_zero <= khugepaged_max_ptes_none) {
1273 result = SCAN_EXCEED_NONE_PTE;
1277 if (!pte_present(pteval)) {
1278 result = SCAN_PTE_NON_PRESENT;
1281 if (pte_uffd_wp(pteval)) {
1283 * Don't collapse the page if any of the small
1284 * PTEs are armed with uffd write protection.
1285 * Here we can also mark the new huge pmd as
1286 * write protected if any of the small ones is
1287 * marked but that could bring uknown
1288 * userfault messages that falls outside of
1289 * the registered range. So, just be simple.
1291 result = SCAN_PTE_UFFD_WP;
1294 if (pte_write(pteval))
1297 page = vm_normal_page(vma, _address, pteval);
1298 if (unlikely(!page)) {
1299 result = SCAN_PAGE_NULL;
1303 if (page_mapcount(page) > 1 &&
1304 ++shared > khugepaged_max_ptes_shared) {
1305 result = SCAN_EXCEED_SHARED_PTE;
1309 page = compound_head(page);
1312 * Record which node the original page is from and save this
1313 * information to khugepaged_node_load[].
1314 * Khupaged will allocate hugepage from the node has the max
1317 node = page_to_nid(page);
1318 if (khugepaged_scan_abort(node)) {
1319 result = SCAN_SCAN_ABORT;
1322 khugepaged_node_load[node]++;
1323 if (!PageLRU(page)) {
1324 result = SCAN_PAGE_LRU;
1327 if (PageLocked(page)) {
1328 result = SCAN_PAGE_LOCK;
1331 if (!PageAnon(page)) {
1332 result = SCAN_PAGE_ANON;
1337 * Check if the page has any GUP (or other external) pins.
1339 * Here the check is racy it may see totmal_mapcount > refcount
1341 * For example, one process with one forked child process.
1342 * The parent has the PMD split due to MADV_DONTNEED, then
1343 * the child is trying unmap the whole PMD, but khugepaged
1344 * may be scanning the parent between the child has
1345 * PageDoubleMap flag cleared and dec the mapcount. So
1346 * khugepaged may see total_mapcount > refcount.
1348 * But such case is ephemeral we could always retry collapse
1349 * later. However it may report false positive if the page
1350 * has excessive GUP pins (i.e. 512). Anyway the same check
1351 * will be done again later the risk seems low.
1353 if (!is_refcount_suitable(page)) {
1354 result = SCAN_PAGE_COUNT;
1357 if (pte_young(pteval) ||
1358 page_is_young(page) || PageReferenced(page) ||
1359 mmu_notifier_test_young(vma->vm_mm, address))
1363 result = SCAN_PAGE_RO;
1364 } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1365 result = SCAN_LACK_REFERENCED_PAGE;
1367 result = SCAN_SUCCEED;
1371 pte_unmap_unlock(pte, ptl);
1373 node = khugepaged_find_target_node();
1374 /* collapse_huge_page will return with the mmap_lock released */
1375 collapse_huge_page(mm, address, hpage, node,
1376 referenced, unmapped);
1379 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1380 none_or_zero, result, unmapped);
1384 static void collect_mm_slot(struct mm_slot *mm_slot)
1386 struct mm_struct *mm = mm_slot->mm;
1388 lockdep_assert_held(&khugepaged_mm_lock);
1390 if (khugepaged_test_exit(mm)) {
1392 hash_del(&mm_slot->hash);
1393 list_del(&mm_slot->mm_node);
1396 * Not strictly needed because the mm exited already.
1398 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1401 /* khugepaged_mm_lock actually not necessary for the below */
1402 free_mm_slot(mm_slot);
1409 * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1410 * khugepaged should try to collapse the page table.
1412 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1415 struct mm_slot *mm_slot;
1417 VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1419 spin_lock(&khugepaged_mm_lock);
1420 mm_slot = get_mm_slot(mm);
1421 if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1422 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1423 spin_unlock(&khugepaged_mm_lock);
1428 * Try to collapse a pte-mapped THP for mm at address haddr.
1430 * This function checks whether all the PTEs in the PMD are pointing to the
1431 * right THP. If so, retract the page table so the THP can refault in with
1434 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1436 unsigned long haddr = addr & HPAGE_PMD_MASK;
1437 struct vm_area_struct *vma = find_vma(mm, haddr);
1439 pte_t *start_pte, *pte;
1445 if (!vma || !vma->vm_file ||
1446 vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE)
1450 * This vm_flags may not have VM_HUGEPAGE if the page was not
1451 * collapsed by this mm. But we can still collapse if the page is
1452 * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1453 * will not fail the vma for missing VM_HUGEPAGE
1455 if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1458 hpage = find_lock_page(vma->vm_file->f_mapping,
1459 linear_page_index(vma, haddr));
1463 if (!PageHead(hpage))
1466 pmd = mm_find_pmd(mm, haddr);
1470 start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1472 /* step 1: check all mapped PTEs are to the right huge page */
1473 for (i = 0, addr = haddr, pte = start_pte;
1474 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1477 /* empty pte, skip */
1481 /* page swapped out, abort */
1482 if (!pte_present(*pte))
1485 page = vm_normal_page(vma, addr, *pte);
1488 * Note that uprobe, debugger, or MAP_PRIVATE may change the
1489 * page table, but the new page will not be a subpage of hpage.
1491 if (hpage + i != page)
1496 /* step 2: adjust rmap */
1497 for (i = 0, addr = haddr, pte = start_pte;
1498 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1503 page = vm_normal_page(vma, addr, *pte);
1504 page_remove_rmap(page, false);
1507 pte_unmap_unlock(start_pte, ptl);
1509 /* step 3: set proper refcount and mm_counters. */
1511 page_ref_sub(hpage, count);
1512 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1515 /* step 4: collapse pmd */
1516 ptl = pmd_lock(vma->vm_mm, pmd);
1517 _pmd = pmdp_collapse_flush(vma, haddr, pmd);
1520 pte_free(mm, pmd_pgtable(_pmd));
1528 pte_unmap_unlock(start_pte, ptl);
1532 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1534 struct mm_struct *mm = mm_slot->mm;
1537 if (likely(mm_slot->nr_pte_mapped_thp == 0))
1540 if (!mmap_write_trylock(mm))
1543 if (unlikely(khugepaged_test_exit(mm)))
1546 for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1547 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1550 mm_slot->nr_pte_mapped_thp = 0;
1551 mmap_write_unlock(mm);
1555 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1557 struct vm_area_struct *vma;
1558 struct mm_struct *mm;
1562 i_mmap_lock_write(mapping);
1563 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1565 * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1566 * got written to. These VMAs are likely not worth investing
1567 * mmap_write_lock(mm) as PMD-mapping is likely to be split
1570 * Not that vma->anon_vma check is racy: it can be set up after
1571 * the check but before we took mmap_lock by the fault path.
1572 * But page lock would prevent establishing any new ptes of the
1573 * page, so we are safe.
1575 * An alternative would be drop the check, but check that page
1576 * table is clear before calling pmdp_collapse_flush() under
1577 * ptl. It has higher chance to recover THP for the VMA, but
1578 * has higher cost too.
1582 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1583 if (addr & ~HPAGE_PMD_MASK)
1585 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1588 pmd = mm_find_pmd(mm, addr);
1592 * We need exclusive mmap_lock to retract page table.
1594 * We use trylock due to lock inversion: we need to acquire
1595 * mmap_lock while holding page lock. Fault path does it in
1596 * reverse order. Trylock is a way to avoid deadlock.
1598 if (mmap_write_trylock(mm)) {
1599 if (!khugepaged_test_exit(mm)) {
1600 spinlock_t *ptl = pmd_lock(mm, pmd);
1601 /* assume page table is clear */
1602 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1605 pte_free(mm, pmd_pgtable(_pmd));
1607 mmap_write_unlock(mm);
1609 /* Try again later */
1610 khugepaged_add_pte_mapped_thp(mm, addr);
1613 i_mmap_unlock_write(mapping);
1617 * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1619 * Basic scheme is simple, details are more complex:
1620 * - allocate and lock a new huge page;
1621 * - scan page cache replacing old pages with the new one
1622 * + swap/gup in pages if necessary;
1624 * + keep old pages around in case rollback is required;
1625 * - if replacing succeeds:
1628 * + unlock huge page;
1629 * - if replacing failed;
1630 * + put all pages back and unfreeze them;
1631 * + restore gaps in the page cache;
1632 * + unlock and free huge page;
1634 static void collapse_file(struct mm_struct *mm,
1635 struct file *file, pgoff_t start,
1636 struct page **hpage, int node)
1638 struct address_space *mapping = file->f_mapping;
1640 struct page *new_page;
1641 pgoff_t index, end = start + HPAGE_PMD_NR;
1642 LIST_HEAD(pagelist);
1643 XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1644 int nr_none = 0, result = SCAN_SUCCEED;
1645 bool is_shmem = shmem_file(file);
1647 VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1648 VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1650 /* Only allocate from the target node */
1651 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1653 new_page = khugepaged_alloc_page(hpage, gfp, node);
1655 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1659 if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1660 result = SCAN_CGROUP_CHARGE_FAIL;
1663 count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1665 /* This will be less messy when we use multi-index entries */
1668 xas_create_range(&xas);
1669 if (!xas_error(&xas))
1671 xas_unlock_irq(&xas);
1672 if (!xas_nomem(&xas, GFP_KERNEL)) {
1678 __SetPageLocked(new_page);
1680 __SetPageSwapBacked(new_page);
1681 new_page->index = start;
1682 new_page->mapping = mapping;
1685 * At this point the new_page is locked and not up-to-date.
1686 * It's safe to insert it into the page cache, because nobody would
1687 * be able to map it or use it in another way until we unlock it.
1690 xas_set(&xas, start);
1691 for (index = start; index < end; index++) {
1692 struct page *page = xas_next(&xas);
1694 VM_BUG_ON(index != xas.xa_index);
1698 * Stop if extent has been truncated or
1699 * hole-punched, and is now completely
1702 if (index == start) {
1703 if (!xas_next_entry(&xas, end - 1)) {
1704 result = SCAN_TRUNCATED;
1707 xas_set(&xas, index);
1709 if (!shmem_charge(mapping->host, 1)) {
1713 xas_store(&xas, new_page);
1718 if (xa_is_value(page) || !PageUptodate(page)) {
1719 xas_unlock_irq(&xas);
1720 /* swap in or instantiate fallocated page */
1721 if (shmem_getpage(mapping->host, index, &page,
1726 } else if (trylock_page(page)) {
1728 xas_unlock_irq(&xas);
1730 result = SCAN_PAGE_LOCK;
1733 } else { /* !is_shmem */
1734 if (!page || xa_is_value(page)) {
1735 xas_unlock_irq(&xas);
1736 page_cache_sync_readahead(mapping, &file->f_ra,
1739 /* drain pagevecs to help isolate_lru_page() */
1741 page = find_lock_page(mapping, index);
1742 if (unlikely(page == NULL)) {
1746 } else if (PageDirty(page)) {
1748 * khugepaged only works on read-only fd,
1749 * so this page is dirty because it hasn't
1750 * been flushed since first write. There
1751 * won't be new dirty pages.
1753 * Trigger async flush here and hope the
1754 * writeback is done when khugepaged
1755 * revisits this page.
1757 * This is a one-off situation. We are not
1758 * forcing writeback in loop.
1760 xas_unlock_irq(&xas);
1761 filemap_flush(mapping);
1764 } else if (PageWriteback(page)) {
1765 xas_unlock_irq(&xas);
1768 } else if (trylock_page(page)) {
1770 xas_unlock_irq(&xas);
1772 result = SCAN_PAGE_LOCK;
1778 * The page must be locked, so we can drop the i_pages lock
1779 * without racing with truncate.
1781 VM_BUG_ON_PAGE(!PageLocked(page), page);
1783 /* make sure the page is up to date */
1784 if (unlikely(!PageUptodate(page))) {
1790 * If file was truncated then extended, or hole-punched, before
1791 * we locked the first page, then a THP might be there already.
1793 if (PageTransCompound(page)) {
1794 result = SCAN_PAGE_COMPOUND;
1798 if (page_mapping(page) != mapping) {
1799 result = SCAN_TRUNCATED;
1803 if (!is_shmem && (PageDirty(page) ||
1804 PageWriteback(page))) {
1806 * khugepaged only works on read-only fd, so this
1807 * page is dirty because it hasn't been flushed
1808 * since first write.
1814 if (isolate_lru_page(page)) {
1815 result = SCAN_DEL_PAGE_LRU;
1819 if (page_has_private(page) &&
1820 !try_to_release_page(page, GFP_KERNEL)) {
1821 result = SCAN_PAGE_HAS_PRIVATE;
1822 putback_lru_page(page);
1826 if (page_mapped(page))
1827 unmap_mapping_pages(mapping, index, 1, false);
1830 xas_set(&xas, index);
1832 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1833 VM_BUG_ON_PAGE(page_mapped(page), page);
1836 * The page is expected to have page_count() == 3:
1837 * - we hold a pin on it;
1838 * - one reference from page cache;
1839 * - one from isolate_lru_page;
1841 if (!page_ref_freeze(page, 3)) {
1842 result = SCAN_PAGE_COUNT;
1843 xas_unlock_irq(&xas);
1844 putback_lru_page(page);
1849 * Add the page to the list to be able to undo the collapse if
1850 * something go wrong.
1852 list_add_tail(&page->lru, &pagelist);
1854 /* Finally, replace with the new page. */
1855 xas_store(&xas, new_page);
1864 __inc_node_page_state(new_page, NR_SHMEM_THPS);
1866 __inc_node_page_state(new_page, NR_FILE_THPS);
1867 filemap_nr_thps_inc(mapping);
1871 __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1873 __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1877 xas_unlock_irq(&xas);
1880 if (result == SCAN_SUCCEED) {
1881 struct page *page, *tmp;
1884 * Replacing old pages with new one has succeeded, now we
1885 * need to copy the content and free the old pages.
1888 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1889 while (index < page->index) {
1890 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1893 copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1895 list_del(&page->lru);
1896 page->mapping = NULL;
1897 page_ref_unfreeze(page, 1);
1898 ClearPageActive(page);
1899 ClearPageUnevictable(page);
1904 while (index < end) {
1905 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1909 SetPageUptodate(new_page);
1910 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1912 set_page_dirty(new_page);
1913 lru_cache_add(new_page);
1916 * Remove pte page tables, so we can re-fault the page as huge.
1918 retract_page_tables(mapping, start);
1921 khugepaged_pages_collapsed++;
1925 /* Something went wrong: roll back page cache changes */
1927 mapping->nrpages -= nr_none;
1930 shmem_uncharge(mapping->host, nr_none);
1932 xas_set(&xas, start);
1933 xas_for_each(&xas, page, end - 1) {
1934 page = list_first_entry_or_null(&pagelist,
1936 if (!page || xas.xa_index < page->index) {
1940 /* Put holes back where they were */
1941 xas_store(&xas, NULL);
1945 VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1947 /* Unfreeze the page. */
1948 list_del(&page->lru);
1949 page_ref_unfreeze(page, 2);
1950 xas_store(&xas, page);
1952 xas_unlock_irq(&xas);
1954 putback_lru_page(page);
1958 xas_unlock_irq(&xas);
1960 new_page->mapping = NULL;
1963 unlock_page(new_page);
1965 VM_BUG_ON(!list_empty(&pagelist));
1966 if (!IS_ERR_OR_NULL(*hpage))
1967 mem_cgroup_uncharge(*hpage);
1968 /* TODO: tracepoints */
1971 static void khugepaged_scan_file(struct mm_struct *mm,
1972 struct file *file, pgoff_t start, struct page **hpage)
1974 struct page *page = NULL;
1975 struct address_space *mapping = file->f_mapping;
1976 XA_STATE(xas, &mapping->i_pages, start);
1978 int node = NUMA_NO_NODE;
1979 int result = SCAN_SUCCEED;
1983 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1985 xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1986 if (xas_retry(&xas, page))
1989 if (xa_is_value(page)) {
1990 if (++swap > khugepaged_max_ptes_swap) {
1991 result = SCAN_EXCEED_SWAP_PTE;
1997 if (PageTransCompound(page)) {
1998 result = SCAN_PAGE_COMPOUND;
2002 node = page_to_nid(page);
2003 if (khugepaged_scan_abort(node)) {
2004 result = SCAN_SCAN_ABORT;
2007 khugepaged_node_load[node]++;
2009 if (!PageLRU(page)) {
2010 result = SCAN_PAGE_LRU;
2014 if (page_count(page) !=
2015 1 + page_mapcount(page) + page_has_private(page)) {
2016 result = SCAN_PAGE_COUNT;
2021 * We probably should check if the page is referenced here, but
2022 * nobody would transfer pte_young() to PageReferenced() for us.
2023 * And rmap walk here is just too costly...
2028 if (need_resched()) {
2035 if (result == SCAN_SUCCEED) {
2036 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2037 result = SCAN_EXCEED_NONE_PTE;
2039 node = khugepaged_find_target_node();
2040 collapse_file(mm, file, start, hpage, node);
2044 /* TODO: tracepoints */
2047 static void khugepaged_scan_file(struct mm_struct *mm,
2048 struct file *file, pgoff_t start, struct page **hpage)
2053 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2059 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2060 struct page **hpage)
2061 __releases(&khugepaged_mm_lock)
2062 __acquires(&khugepaged_mm_lock)
2064 struct mm_slot *mm_slot;
2065 struct mm_struct *mm;
2066 struct vm_area_struct *vma;
2070 lockdep_assert_held(&khugepaged_mm_lock);
2072 if (khugepaged_scan.mm_slot)
2073 mm_slot = khugepaged_scan.mm_slot;
2075 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2076 struct mm_slot, mm_node);
2077 khugepaged_scan.address = 0;
2078 khugepaged_scan.mm_slot = mm_slot;
2080 spin_unlock(&khugepaged_mm_lock);
2081 khugepaged_collapse_pte_mapped_thps(mm_slot);
2085 * Don't wait for semaphore (to avoid long wait times). Just move to
2086 * the next mm on the list.
2089 if (unlikely(!mmap_read_trylock(mm)))
2090 goto breakouterloop_mmap_lock;
2091 if (likely(!khugepaged_test_exit(mm)))
2092 vma = find_vma(mm, khugepaged_scan.address);
2095 for (; vma; vma = vma->vm_next) {
2096 unsigned long hstart, hend;
2099 if (unlikely(khugepaged_test_exit(mm))) {
2103 if (!hugepage_vma_check(vma, vma->vm_flags)) {
2108 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2109 hend = vma->vm_end & HPAGE_PMD_MASK;
2112 if (khugepaged_scan.address > hend)
2114 if (khugepaged_scan.address < hstart)
2115 khugepaged_scan.address = hstart;
2116 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2117 if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
2120 while (khugepaged_scan.address < hend) {
2123 if (unlikely(khugepaged_test_exit(mm)))
2124 goto breakouterloop;
2126 VM_BUG_ON(khugepaged_scan.address < hstart ||
2127 khugepaged_scan.address + HPAGE_PMD_SIZE >
2129 if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2130 struct file *file = get_file(vma->vm_file);
2131 pgoff_t pgoff = linear_page_index(vma,
2132 khugepaged_scan.address);
2134 mmap_read_unlock(mm);
2136 khugepaged_scan_file(mm, file, pgoff, hpage);
2139 ret = khugepaged_scan_pmd(mm, vma,
2140 khugepaged_scan.address,
2143 /* move to next address */
2144 khugepaged_scan.address += HPAGE_PMD_SIZE;
2145 progress += HPAGE_PMD_NR;
2147 /* we released mmap_lock so break loop */
2148 goto breakouterloop_mmap_lock;
2149 if (progress >= pages)
2150 goto breakouterloop;
2154 mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2155 breakouterloop_mmap_lock:
2157 spin_lock(&khugepaged_mm_lock);
2158 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2160 * Release the current mm_slot if this mm is about to die, or
2161 * if we scanned all vmas of this mm.
2163 if (khugepaged_test_exit(mm) || !vma) {
2165 * Make sure that if mm_users is reaching zero while
2166 * khugepaged runs here, khugepaged_exit will find
2167 * mm_slot not pointing to the exiting mm.
2169 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2170 khugepaged_scan.mm_slot = list_entry(
2171 mm_slot->mm_node.next,
2172 struct mm_slot, mm_node);
2173 khugepaged_scan.address = 0;
2175 khugepaged_scan.mm_slot = NULL;
2176 khugepaged_full_scans++;
2179 collect_mm_slot(mm_slot);
2185 static int khugepaged_has_work(void)
2187 return !list_empty(&khugepaged_scan.mm_head) &&
2188 khugepaged_enabled();
2191 static int khugepaged_wait_event(void)
2193 return !list_empty(&khugepaged_scan.mm_head) ||
2194 kthread_should_stop();
2197 static void khugepaged_do_scan(void)
2199 struct page *hpage = NULL;
2200 unsigned int progress = 0, pass_through_head = 0;
2201 unsigned int pages = khugepaged_pages_to_scan;
2204 barrier(); /* write khugepaged_pages_to_scan to local stack */
2206 lru_add_drain_all();
2208 while (progress < pages) {
2209 if (!khugepaged_prealloc_page(&hpage, &wait))
2214 if (unlikely(kthread_should_stop() || try_to_freeze()))
2217 spin_lock(&khugepaged_mm_lock);
2218 if (!khugepaged_scan.mm_slot)
2219 pass_through_head++;
2220 if (khugepaged_has_work() &&
2221 pass_through_head < 2)
2222 progress += khugepaged_scan_mm_slot(pages - progress,
2226 spin_unlock(&khugepaged_mm_lock);
2229 if (!IS_ERR_OR_NULL(hpage))
2233 static bool khugepaged_should_wakeup(void)
2235 return kthread_should_stop() ||
2236 time_after_eq(jiffies, khugepaged_sleep_expire);
2239 static void khugepaged_wait_work(void)
2241 if (khugepaged_has_work()) {
2242 const unsigned long scan_sleep_jiffies =
2243 msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2245 if (!scan_sleep_jiffies)
2248 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2249 wait_event_freezable_timeout(khugepaged_wait,
2250 khugepaged_should_wakeup(),
2251 scan_sleep_jiffies);
2255 if (khugepaged_enabled())
2256 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2259 static int khugepaged(void *none)
2261 struct mm_slot *mm_slot;
2264 set_user_nice(current, MAX_NICE);
2266 while (!kthread_should_stop()) {
2267 khugepaged_do_scan();
2268 khugepaged_wait_work();
2271 spin_lock(&khugepaged_mm_lock);
2272 mm_slot = khugepaged_scan.mm_slot;
2273 khugepaged_scan.mm_slot = NULL;
2275 collect_mm_slot(mm_slot);
2276 spin_unlock(&khugepaged_mm_lock);
2280 static void set_recommended_min_free_kbytes(void)
2284 unsigned long recommended_min;
2286 for_each_populated_zone(zone) {
2288 * We don't need to worry about fragmentation of
2289 * ZONE_MOVABLE since it only has movable pages.
2291 if (zone_idx(zone) > gfp_zone(GFP_USER))
2297 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2298 recommended_min = pageblock_nr_pages * nr_zones * 2;
2301 * Make sure that on average at least two pageblocks are almost free
2302 * of another type, one for a migratetype to fall back to and a
2303 * second to avoid subsequent fallbacks of other types There are 3
2304 * MIGRATE_TYPES we care about.
2306 recommended_min += pageblock_nr_pages * nr_zones *
2307 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2309 /* don't ever allow to reserve more than 5% of the lowmem */
2310 recommended_min = min(recommended_min,
2311 (unsigned long) nr_free_buffer_pages() / 20);
2312 recommended_min <<= (PAGE_SHIFT-10);
2314 if (recommended_min > min_free_kbytes) {
2315 if (user_min_free_kbytes >= 0)
2316 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2317 min_free_kbytes, recommended_min);
2319 min_free_kbytes = recommended_min;
2321 setup_per_zone_wmarks();
2324 int start_stop_khugepaged(void)
2328 mutex_lock(&khugepaged_mutex);
2329 if (khugepaged_enabled()) {
2330 if (!khugepaged_thread)
2331 khugepaged_thread = kthread_run(khugepaged, NULL,
2333 if (IS_ERR(khugepaged_thread)) {
2334 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2335 err = PTR_ERR(khugepaged_thread);
2336 khugepaged_thread = NULL;
2340 if (!list_empty(&khugepaged_scan.mm_head))
2341 wake_up_interruptible(&khugepaged_wait);
2343 set_recommended_min_free_kbytes();
2344 } else if (khugepaged_thread) {
2345 kthread_stop(khugepaged_thread);
2346 khugepaged_thread = NULL;
2349 mutex_unlock(&khugepaged_mutex);
2353 void khugepaged_min_free_kbytes_update(void)
2355 mutex_lock(&khugepaged_mutex);
2356 if (khugepaged_enabled() && khugepaged_thread)
2357 set_recommended_min_free_kbytes();
2358 mutex_unlock(&khugepaged_mutex);